Documentation officielle de Google pour le prompt dans Gemini

Décortiquer comment Google AI Overview construit ses réponses.

1) RAG Implementation Exposed: The author discovered a prompt technique that reveals the exact data Google feeds to Gemini from its search index. Each result contains title, URL, snippet, and an "additional_info" field that appears to be AI-generated summarization rather than direct website content.

2) JSON Structure: When Gemini processes a query, it receives search results in a specific JSON format with fields including "type," "source," "title," "url," "snippet," and "additional_info." This grounding context enables the model to access current information beyond its training data.

3) Context Size Variation: While most grounding contexts contain only short snippets, the author discovered that in some cases, Google provides multi-paragraph snippets to Gemini, giving the AI more comprehensive context for certain queries.

4) Dynamic Retrieval Mechanism: Google employs a "confidence score" system to determine when grounding is needed. If the model's prediction score exceeds a threshold (default 0.3), the response is grounded with search results; otherwise, the model answers without grounding, potentially causing hallucinations.

5) Brand Representation Impact: This revelation highlights the importance of SEO in AI-driven discovery, as the snippets Google selects for grounding directly influence how brands are represented to users in AI responses. The author suggests using a tool called AI Rank to track how brands are associated with relevant queries.